[Issue] Linux PyTorch nightly tests fail across six GPU families: pytest rejects --exclude-jit-executor #9899

Context

• Workflow: multi_arch_release_linux_pytorch_wheels.yml (ROCm/rockrel)
• Failing run: 42970099850, attempt 1
• Platform: Linux
• Affected architectures: gfx94X-dcgpu, gfx120X-all, gfx110X-all, gfx950-dcgpu, gfx90a, gfx1151
• PyTorch channel: nightly
• Python version: 3.11
• ROCm package version: 10.3.0a20261111

Summary

All six failed Test PyTorch jobs stop before test collection because pytest rejects --exclude-jit-executor. The error occurs in test_autograd.py, test_binary_ufuncs.py, test_cuda.py, test_nn.py, test_torch.py, and test_unary_ufuncs.py across these architectures. The missing JUnit XML errors follow the argument parsing failures; no individual test assertion is reached.

TheRock's external-builds/pytorch/run_pytorch_tests_full.py enables and forwards this option by default. The observed nightly test runner passes it to individual pytest commands, which reject it. The logs also warn that skip_tests/pytorch_2.16.py is absent, but the argument error is the immediate test blocker.

Exact Error

From the gfx110X job:

2026-11-11T10:04:53.3585286Z UsageError: usage: test_autograd.py [options] [file_or_dir] [file_or_dir] [...]
2026-11-11T10:04:53.3585930Z test_autograd.py: error: unrecognized arguments: --exclude-jit-executor
2026-11-11T10:04:53.3588380Z test_autograd.py: error: unrecognized arguments: --exclude-jit-executor
2026-11-11T10:04:53.3589536Z Test results will be stored in test-reports/python-pytest/test_autograd/test_autograd-c9b01f0a86343064.xml
2026-11-11T10:04:53.3598978Z FileNotFoundError: [Errno 2] No such file or directory: 'test-reports/python-pytest/test_autograd/test_autograd-c9b01f0a86343064.xml'

Failing jobs

Root cause commit

The jobs pin the wheel/test source to ROCm/pytorch@df21fc2d. ROCm/pytorch@5a989117 removed --exclude-jit-executor from test/run_test.py's argument parser and deleted the associated JIT executor test exclusions. The preceding nightly, 75b45b7c0c7409a0374c4a89a4a4549bbddf6af1, still accepts the option, while df21fc2d does not. The change originates in upstream pytorch/pytorch@6dca9f63 (#203621). TheRock's existing wrapper still sends the removed flag by default.

Proposed fix

Update TheRock's run_pytorch_tests_full.py to omit --exclude-jit-executor for PyTorch versions containing the removal and update its wrapper option/help text. If supported older release branches still need the exclusion, gate it by PyTorch version. A short-term validation is to run one failing module with the wrapper's --no-exclude-jit-executor option, then rerun the affected GPU jobs. This fix is proposed; it has not been tested here.